Saigar’s Desk
A personal research operator, in the principal’s voice.
I’ve spent the last few months building Saigar, a multi-agent research system that runs financial-sector research pipelines on my behalf and produces artefacts to my standards. The problem it solves is the one I run into every week: reading is unbounded, writing is bounded by my time, and the synthesis between them is what’s actually scarce.
Eighteen agents do the work. Eight are senior generalists that show up across every pipeline: an orchestrator that sequences stages, a researcher that gathers and triages source material, a synthesizer that turns evidence into structured claim and evidence cards, a writer that drafts prose in voice, a critic that runs voice rules and peer review, a reflector that writes lessons back into a learning-signals table after each run, a designer that lays briefs out into the desk’s paper archetypes, and a publisher that commits everything to the live newspaper. Ten more are junior specialists. Three run on the wire pipeline (a curator, a drafter, and a critic, each tuned to the 95 to 110 word format that lives in the briefing column). Seven run on the company-assessment pipeline: a company researcher that pulls filings and IR feeds for one specific company, a metrics extractor that lifts a standard KPI schema from those filings, a bear-case analyst and a bull-case analyst that argue opposite sides as falsifiable predictions, a reconciler that adjudicates the red-team into a single house view, a ratification agent that scores the prior edition’s predictions against what actually happened, and a title generator.
The system runs four pipelines today, in sixty-six total stages. Briefs are short weekly takes on a topic in motion. The writer drafts, the critic enforces voice, the designer lays them out as a desk-paper artefact, and the publisher ships them. Working papers are longer, figure-heavy pieces that go through a deeper research and synthesis loop, with the reflector feeding lessons back into the learning-signals table so the desk gradually inherits my taste in framing, sourcing, and structure. Wire snippetsship at 07:00 CEST every morning: the curator picks up to five items from the last 24 hours of regulator filings, central-bank data releases, working papers, and payments-domain news, the drafter writes a 95 to 110 word note for each, the critic scans for voice and citation discipline, and the publisher commits them to the “On the wire” column at desk.sagarbharambe.com. One source per snippet, dated, capped to one per channel per day. Predictive Intelis the newest pipeline: quarterly company assessments for seven payment companies (Adyen, Mastercard, PayPal, Stripe, Visa, Wise, Worldline). A bear analyst and a bull analyst each generate falsifiable predictions in their own voice. A reconciler reads both and writes the house view as a single integrated draft, surfacing the predictions where the two disagreed. From edition two onward, a ratification scorecard opens the report and grades the prior edition’s predictions against what actually happened. The first edition (PayPal, May 2026) is live at desk.sagarbharambe.com/predictive-intel.
The Brief 47 example threading through the showcase below is a real working subject: agentic commerce identity infrastructure, with Mastercard Agent Pay, Visa Intelligent Commerce, the platform protocols (ACP and UCP), and Web Bot Auth as the moving pieces.
The mechanism diagram animates. The run simulator on the Desk page plays through a complete cycle. The newspaper and slides toggle in the masthead is a meta-demonstration of one of Saigar’s actual product features. The whole thing is a single self-contained HTML artefact, around 180 KB, running entirely in the browser.
↓ Interactive showcase below. Don’t be afraid to click around.